Development and Validation of a Deep Learning Based Diabetes Prediction System Using a Nationwide Population-Based...

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Epidemiology

Background:

• Existing type 2 diabetes mellitus (T2DM) prediction models demonstrate limited efficacy.
• A novel deep learning (DL) approach was developed for T2DM risk prediction.
• The model was trained and validated on a large cohort representative of the Korean population.

Purpose of the Study:

• To develop and evaluate a DL-based model for predicting T2DM.
• To compare the performance of the DL model against a conventional Cox model.
• To identify key risk factors for T2DM using advanced predictive modeling.

Main Methods:

• Utilized the National Health Insurance Service-Health Screening (NHIS-HEALS) cohort of 335,302 Korean subjects.
• Developed a recurrent neural network long short-term memory (RNN-LSTM) model for T2DM prediction.
• Compared the DL model's 10-year predictive performance against a Cox longitudinal summary model using time-dependent AUC.

Main Results:

• The RNN-LSTM DL model exhibited superior annual performance compared to the Cox model over a 10-year follow-up period.
• T2DM was newly diagnosed in 8.7% of subjects during the mean 10.4-year follow-up.
• Identified similar, yet distinct, risk factors for T2DM between the DL and Cox models.

Conclusions:

• The DL-based T2DM prediction model offers improved performance over conventional methods for the Korean population.
• The model is slated for future implementation in national health screening programs following pilot testing.
• This study highlights the potential of DL in enhancing T2DM risk prediction and management.